Four ways to work with me.
Every engagement is remote-first, scoped in writing before any work starts, and sized so you own the result — with me or without me. I take a limited number of engagements: one at a time, done properly.
Start with the assessment. It's fixed-price, it's useful on its own, and it tells us both whether the bigger work makes sense.
Sovereignty Assessment — €3,500, fixed.
You want AI, your data can't leave, and you need to know what's actually possible before you spend real money. That's this.
Over one to two weeks, fully remote, I review:
- · What your team wants from AI — and what it genuinely needs
- · Your data flows and where the sensitivity actually lives
- · Your compliance constraints (GDPR, NIS2, DORA, sector rules)
- · Your existing infrastructure, and what hardware you'd need
You get a written architecture recommendation:
- · The data boundary, drawn explicitly — what must stay on local models, what may route to frontier models, and how the router enforces the line so it's policy, not promises
- · Hardware sizing and procurement guidance — the exact class of machine, realistic prices, what to avoid
- · A cost projection built on real per-call data — I run this stack; my numbers come from dashboards, not vendor decks
- · A concrete deployment proposal — scope, timeline, fixed quote
If the honest answer is “self-hosted AI isn't right for you yet,” the report says so and explains why. You'll have paid €3,500 to not waste €50,000. That's a good trade.
Foundation Deployment — your stack, your hardware, your keys.
The production build. On your servers, in your building (or your EU datacenter), I deploy:
- · A chat workspace your team actually uses — clean UI, agents, file handling, search. Built on proven open source, tracking upstream — no fork you're trapped in.
- · One router in control of every AI call — local models and commercial providers behind a single API. Swap providers with a config change, never an application rewrite. Per-call cost tracking. Budgets per team, per key.
- · The data boundary, enforced — sensitive routes pinned to local models on your hardware. Commodity work goes where you allow it. Every call logged; every euro visible.
- · Monitoring from day one — dashboards for health, usage, and spend. Your finance team sees costs; your security team sees flows.
- · Backups and recovery, tested — not a checkbox. A restore procedure your team has actually seen run.
And the part most vendors skip: the handover. Documentation written to be operated by someone who isn't me — runbooks, decision records, a training session for your operator. The day I hand it over, it's yours. No retainer required, no dependency by design.
Typical range: €5,000–10,000, fixed quote after assessment. Timeline: weeks, not quarters.
Document Intelligence — your documents, readable by AI, without leaving the building.
Most teams discover the same thing right after deploying private AI: the value is locked in documents. Contracts, dossiers, reports, scans — and every cloud OCR API is a data-export event you can't sign off.
I build a local document pipeline that:
- · Routes each page to the right engine — born-digital text is extracted directly (exact, instant), scanned pages go to OCR, and pages with embedded diagrams go to a local vision model that can actually read them
- · Never lies about quality — every run produces a manifest: pages that are clean, pages with benign warnings, and the short list that genuinely needs human eyes. Silent failure is designed out, not hoped away.
- · Versions everything — new revision of a document? It lands next to the old one, diffable, with the current version always addressable. Identical re-drops are skipped automatically.
- · Runs itself — drop files in a folder; the pipeline picks them up, processes them in order, and notifies you when done.
Available as an add-on to a Foundation Deployment or standalone if you already run your own AI stack. Scoped per engagement.
Assessment Workflows — AI that reviews documents against your rules, and can't cheat.
For teams whose work is judgment at scale: security architecture reviews, supplier due diligence, DORA/NIS2 readiness, contract review against playbooks. AI can carry real weight here — if, and only if, it's built so the two failure modes that destroy trust are impossible:
Made-up findings and silently skipped checks.
I design assessment workflows where honesty is enforced by code, not by prompt:
- · A maker–checker agent design — one accountable lead agent that assesses under an immutable rule contract; one narrow validator that can only verify, never invent; one scribe that renders the discussion into human-readable records and is forbidden to judge
- · A deterministic verifier between the AI and you — every finding must cite a source that exists, on a page that exists, with wording that's actually there. Every in-scope check must be accounted for — assessed or explicitly flagged, never quietly dropped. This layer is code. It doesn't have good days and bad days.
- · Your framework, not mine — your control set, your playbook, your review rules, frozen at engagement start so results are reproducible against a known baseline
- · The human stays in charge — the system raises and evidences; your experts rule. Always. A workflow that pre-decides for your reviewers is a workflow your reviewers will rightly reject.
This is the deepest work I do, and it's scoped individually — typically after a Foundation Deployment, sometimes on top of an existing stack.
One honest prerequisite: real hardware.
Serious local AI — the kind that reads your documents with 30-billion-parameter models and answers in seconds — needs a serious machine: roughly 96–128 GB of unified memory and a capable GPU. Today that's a €2,500–4,000 investment, made once, owned by you.
If that's not in your budget, self-hosted AI is not your path yet — and I'll tell you so rather than deploy something that disappoints. The assessment includes exact sizing and procurement guidance, so if you do buy, you buy right the first time.
What's deliberately not on offer.
- No retainer requirement. Handover means handover. If you want ongoing operations support afterwards, we can talk — but it's never a condition.
- No per-seat, per-month meter. You buy builds, not subscriptions.
- No lock-in. Everything I deploy, you can run without me the day it's delivered. That's checked, not claimed — your operator runs the runbook before I leave.
- No overselling. If a ChatGPT Team plan honestly covers your need, that's what I'll tell you. It's cheaper for you and better for my sleep.
Tell me what you're trying to do.
One email. I'll tell you honestly whether I can help — and if I can't, I'll point you somewhere that can. I reply within one business day.
Talk to me — I don't biteScoped per engagement · Remote-first, EU · One engagement at a time